Annotation Versioning and Cross-Collection Generalization in Varroa Mite Counting

Reliable Varroa mite counting requires evaluation protocols that distinguish annotation effects, shared image content, and transfer across acquisition sources. This study examines these factors through a paired annotation audit, spatial-overlap screening, inference controls, and a frozen-checkpoint evaluation of a YOLO11n detector on a second public collection. The first dataset contains 64 photographs and two annotation representations: 807 CSV boxes and 849 XML boxes. All 42 additional XML boxes occur in the 14-image validation partition. Holding the model, predictions, and operating threshold fixed, changing the reference from CSV to XML changes AP50 from 0.8981 to 0.9130, precision from 0.7345 to 0.8276, and mean absolute count error from 4.64 to 3.36 per photograph. Geometric matching and visual inspection identify shared scene regions in three training-validation image pairs. An additional collection contains 180 crops with 8,572 annotations and 25 filename-derived parent groups; 22 groups span the uploader's partitions. On this collection, the unchanged detector produces no detections at the prespecified confidence threshold of 0.25. A known positive source tile reproduces its cached confidence exactly, providing a runtime control. The contribution is a reproducible empirical account of how annotation representation, sampling units, and collection transfer constrain the interpretation of a detection benchmark. The findings apply to the evaluated checkpoint and releases; independent field performance remains unestablished. The deposit includes the preprint and a companion reproducibility package with versioned references, frozen predictions, audit records, three portable Python scripts, and quantitative figures. All eight metric configurations were checked against saved values. Same-runtime inference reruns reproduced 1,245 source and 921 additional-collection prediction records exactly. Source validation was used for checkpoint selection; the additional collection has no empty crops and uncertain upstream acquisition identities. Acknowledgments: EVREN and Türkiye's Presidency of Defence Industries (SSB) provided the platform used for model training and sharing. Data preparation, provenance auditing, and ONNX evaluation ran locally. The author thanks Jose Divasón and colleagues for the sticky-board collection (https://doi.org/10.5281/zenodo.10231845), Tomislav uri for the additional collection (https://universe.roboflow.com/tomislav-uri-lv8pr/varroa-tdfik/dataset/4), and the open-source tool maintainers. These acknowledgments do not imply sponsorship or endorsement. Public trained model, v1.0.0: https://evren.ssyz.org.tr/models/uda23e07a83e0/varroa-tile1024-yolo11n-baseline . The model card specifies AGPL-3.0; weights are not included in this deposit. Source annotations and attributed image previews retain CC BY 4.0 terms. Newly authored portable scripts use the included MIT license; other file-specific rights are documented in ATTRIBUTION.md. OpenAI Codex assisted with literature retrieval, source inspection, code generation and execution, analysis, image inspection, and manuscript drafting. Reported measurements were computed from the stated files and model inference. This preprint has not been peer reviewed.